Army Soldier Leaked CEO Call Logs
Also Cheap AI and a Web Summarizer Nobody Asked For

The IT Guy Who Went Rogue on the Company Phone Bill

The national-scale data breach at two of the country's largest phone carriers, AT&T and Verizon, has been downgraded from a sophisticated foreign threat to just another HR problem. Federal authorities arrested U.S. Army Soldier Cameron John Wagenius, who operated under the alias "Kiberphant0m" and was charged with two counts of unlawful transfer of confidential phone records. This individual, a communications specialist stationed in South Korea, allegedly extorted the companies after stealing sensitive customer call logs and other proprietary data from a breach that affected over 110 million people.

The plot thickens with the most mundane of corporate security oversights, as Mr. Wagenius is linked to the broader, embarrassing Snowflake hacking campaign that successfully compromised dozens of companies. The data exfiltration was not the result of a zero-day exploit or even a particularly clever virus, but largely due to companies not enforcing basic multi-factor authentication on cloud storage accounts, a lapse equivalent to leaving the server room key under the doormat. The arrested soldier reportedly even tried to sell a SIM-swapping service on Telegram; just another example of an overworked systems administrator deciding that the most efficient path to a raise is by holding the executives' private conversations for ransom.

Mozilla Launches a Product Only Mozilla Could Love

Mozilla, the organization that manages the Firefox web browser, has officially entered the AI space with a product named Orbit. Orbit is a browser extension that summarizes emails, long web pages, and video transcripts, but its real hook is its staunch commitment to privacy. The company is running the Mistral 7B large language model on its own servers and states that it does not save session data or use user queries to train the model, essentially creating an AI that is contractually obligated to forget every conversation you have with it.

The introduction even includes a satirical, passive-aggressive email mocking the aggressive marketing of other AI tools, a move that perfectly captures the company's slightly exhausted position in the tech ecosystem. Launching a sophisticated, self-hosted, privacy-first AI summarizer feels like building a beautiful, handcrafted mahogany desk when the entire office already uses cheap, mass-produced plastic trays, but at least the data remains on a server that is definitely not listening to you.

Costco AI Hits the American Stock Market

A Chinese AI company, Deepseek, caused a significant market flinch after releasing a large language model that reportedly cost a tiny fraction of what American giants like OpenAI spend. The model, R1, managed to achieve impressive performance on reasoning and math metrics for a purported cost of around $6 million for the final training run, compared to the estimated $100 million or more spent by its heavily funded Western rivals. The breakthrough largely relied on efficiency techniques, specifically pioneering a method called sparsity, which lets the model use less computational power per query.

The entire premise of the AI arms race rested on the idea that success was proportional to the amount of money you could throw at expensive Nvidia chips. Deepseek's arrival suggests that you can, in fact, build a world-class model without buying every single H100 GPU on the planet, causing a predictable wave of panic on Wall Street. The new narrative is that the price of doing world-class AI just dropped, which means all the companies that based their billion-dollar valuation on requiring massive expenditure may have a very awkward board meeting scheduled.

Briefs

• 400 Years of Bronze: Zildjian, the Massachusetts-based company that has been making cymbals since the Ottoman Empire, continues to be a successful business, proving that slow, deliberate, physical-world excellence remains a viable business model. Maybe all that VC money should just be spent on metal alloys.
• The Retro Port That Was Not Enough: The fully native Grand Theft Auto III port for the Sega Dreamcast has been officially released. This is proof that a lot of developers would rather spend an unreasonable amount of time porting a twenty-year-old game to a dead console than start a greenfield project in a new language.
• LLM Takeaway: A comprehensive review of LLM developments in 2024 concluded that the systems somehow got even harder to use effectively. Turns out giving a computer the ability to understand all human knowledge does not automatically make it intuitive; it just means the documentation is now 10,000 pages longer.